Gadget of the Week: An AI that
thinks before it speaks

What is it?

Ask a roomful of South Africans which AI chatbot lives on their phone, and almost all will say ChatGPT. Increasingly, however, they will mention Claude as one they’ve heard about but have yet to try out.

Claude’s parent company Anthropic is in the news for both the right and wrong reasons, with the US government banning the use by foreign nationals of its latest AI large language model, Fable 5, because it is regarded as being too powerful to fall into the “wrong” hands.

But even Claude, available in both free and paid versions in South Africa, is a cut above the rest, for one simple reason: It “thinks” before it speaks. Or, at least, it asks clarifying questions before it hazards a guess.

I have used a paid version of Claude for several months, mostly the current Sonnet 4.6 model running inside the ordinary Claude.ai web and desktop apps, for research, editing, document analysis and one deliberately unfair test: asking it to write a press release that did not read like a press release.

If you feed Claude a brief and ask it to avoid the stock phrases of machine-written content – the historic scene-setting (“in a world where…”), the lists of three, the negations (“it’s not this, it’s that”) – and it largely complies, because Anthropic has clearly put real effort into training that tone out of its own model. Whether the same discipline survives in the cheaper tools built on top of Claude’s underlying technology, churning out marketing copy by the thousand, is a different and less reassuring question.

I asked Claude to review itself, and this it told me:

“The product itself follows a familiar shape: a chat window across web, desktop and mobile, a coding-focused sibling called Claude Code for developers, document and spreadsheet tools that plug into Word, Excel and PowerPoint, a memory feature that recalls earlier conversations, and “Projects” that group related chats and files so a long piece of research does not dissolve into a hundred disconnected threads.

“None of this is exclusive to Claude. Google’s Gemini does much the same from inside Workspace and Android, and Microsoft’s Copilot does it from inside Office, software most South African companies already pay for. What separates Claude is less the toolset than the model’s manner: longer, more structured answers, a habit of asking before assuming, and a visible reluctance to bluff when it is unsure, although that same reluctance occasionally tips into refusing requests that are perfectly reasonable.

“On accuracy, Claude is no more immune to being wrong with a straight face than ChatGPT or Gemini. Ask it about something outside its training, or about events from the past week, and it now searches the web rather than guessing, an improvement on the older habit of inventing a plausible answer from nothing.

“Push it on a niche South African fact, though, such as the precise attendance at a regional conference, and it will occasionally produce a number that sounds correct and isn’t, delivered in the same calm, declarative prose it uses for facts it has actually checked. That calm is the catch: Claude rarely hedges, which makes its rare wrong answers harder to spot than those of a chatbot that hedges constantly.”

Such refreshing honesty is in fact a hallmark of Claude. And it undersells itself in this regard. I initially took the Anthropic plunge because I was frustrated that both Gemini and ChatGPT insisted on inventing quotes when I asked it to look for expert analyst comment on a new enterprise software tool.  When I asked Claude to do the same, it told me that no such quotes existed, because the product was too new to have garnered analyst assessments.

One major drawback, however, is the rate at which it uses up “tokens”, a kind of currency that dictates how much content Claude can take in and put out. For content-intensive projects, the tokens run out fast and one has to pay in extra to top up the tokens if in the middle of an urgent project.

That, by the way, was one of the three biggest negatives Claude itself gave me for itself (see below).

Such “honesty” speaks not only of a reliable and consistent AI assistant, but also one that is not going to keep you running round in circles trying to find verified information. In the middle of 2026, with the generative AI era not yet 4-years old, that approach represents remarkable design maturity.

How much does it cost?

The free tier of Claude costs nothing but limits how much can be asked of it in a day, and excludes Claude Code entirely. Claude Pro runs to $20 a month, paid monthly, which at current exchange rates is roughly R330, or about R279 a month if paid upfront for the year at a discounted $17-a-month annual rate. Heavier users may go for Claude Max, from $100 to $200 a month, somewhere between R1 640 and R3 280. South Africa has not been added to the list of countries where Anthropic offers app-store pricing adjusted for local buying power, the way it has in Nigeria or Turkey, so local subscribers pay close to the full dollar price, plus whatever foreign-transaction loading their bank adds on top. Claude is available here through the website, and iOS and Android stores for mobile apps.

Does it make a difference?

For people doing serious writing, editing or analytical work, the longer, more deliberate answers and the document-handling tools save real time once the workflow is set up. For the much larger group of South Africans who simply want quick answers on a phone they already own, the difference is harder to argue, given that Gemini comes free on most Android handsets and Copilot is increasingly bundled into Office subscriptions companies are paying for regardless. Claude’s case rests on being chosen on purpose, as a product that is building a massive word-of-mouth reputation.

What are the biggest negatives?

• The subscription is billed in dollars with no local pricing adjustment, making it one of the costlier AI tools available to South African consumers.
• Heavy users on the Pro plan hit usage limits inside a rolling five-hour window that resets at an inconvenient time more often than it should.

What are the biggest positives?

• Document analysis across PDFs, spreadsheets and slide decks is fast and accurate in most tests, saving real time on research-heavy work.
• The model is more willing to say it does not know something than to bluff.

* Arthur Goldstuck is CEO of World Wide Worx, editor-in-chief of Gadget.co.za, and author of “The Hitchhiker’s Guide to AI – The African Edge”.